Short answer: to find out whether ChatGPT cites your website, ask it a set of questions your pages should be the natural source for, without naming your brand, and check whether your domain appears among the sources marked in the answer. Run each question more than once, because answers vary. If your domain never appears, the cause is usually one of three things: ChatGPT cannot reach or read the page, the page has no passage worth lifting, or other sites carry the facts and you do not. Each needs a different fix.
You can do the whole check in an afternoon with a spreadsheet. The useful part is not the yes or no; it is knowing which of those failures a "no" points to. This article covers the access check, a test procedure that produces evidence rather than anecdotes, how to read each kind of negative result, and the counting mistakes that make automated checks lie.
First, make sure ChatGPT can reach your website
OpenAI runs separate agents for separate jobs, and only some of them matter for citations. According to OpenAI's bot documentation, OAI-SearchBot builds the index that surfaces websites in ChatGPT's search features, ChatGPT-User fetches a page because a person in ChatGPT or a Custom GPT asked for it, and GPTBot crawls content that may be used to train OpenAI's foundation models.
| Agent | Job | Blocking it affects citations? |
|---|---|---|
OAI-SearchBot | Builds the index behind ChatGPT's search features | Yes: you are removed from ChatGPT search results |
ChatGPT-User | Fetches a page a user asked about | Yes: a user who asks ChatGPT to read your page gets nothing |
GPTBot | Collects content that may be used for training | No effect on today's answers |
One rule that blocks all three is the most common way sites remove themselves from ChatGPT's answers, and OAI-SearchBot is the agent most often blocked by accident by sites that meant to opt out of training. If declining training is your goal, block GPTBot by name and leave the other two alone.
The four checkpoints before a citation is possible
robots.txt is only the first gate. A page has to pass four checkpoints before ChatGPT can cite it, and the last three never show up in the file:
- robots.txt allows OAI-SearchBot. Read the group that names it, or the
*group if none does. A named group does not inherit the rules written under*. - The edge returns the page. Request a page with the OAI-SearchBot user agent and compare the response with a normal browser request. A CDN or firewall rule that returns 403 to the agent blocks it in practice, while robots.txt keeps saying otherwise.
- The content is in the HTML. Most AI crawlers do not execute JavaScript, so a page whose text appears only after client-side rendering is an empty shell to them. A 200 response with an empty body is a failure, not a pass.
- A passage answers the question. If the first two sentences under a heading do not answer that heading's question on their own, there is little for an answer to lift.
Run the probe from a second network as well, and keep a plain browser request beside every agent request. Without that control, a site that is briefly down looks exactly like a site that blocks AI agents. Our AI crawler access guide covers the same checks for the other engines.
How to check whether ChatGPT cites your website
A citation check is easy to do in a way that proves nothing. Four rules prevent that:
- Use questions that do not contain your brand name. A question that names you retrieves your site trivially. Ask what a buyer would ask before they know you exist: "which accounting software handles multi-currency invoices for a small agency", not "is Acme good". In one of our own early prompt sets, 13 of 15 generated questions carried the brand name, and the resulting score was meaningless.
- Write full situations, not keywords. Assistants answer situations. A question carrying a budget, a constraint or a use case produces a different set of sources than the bare category term, and it is closer to what people actually type.
- Run each question more than once. Answers vary between runs even when nothing has changed. One run is an anecdote; the same result across runs is evidence.
- Count only sources marked in the answer. Your domain counts when it appears as a source attached to the answer text, and only if the marker resolves to a real URL.
Make the set large enough that one odd answer cannot swing the result, and spread it across the jobs your pages do: explaining a concept, comparing options, answering a specific constraint. Building a prompt set covers how to choose them and how to keep the set fixed so results stay comparable.
Record every run in the same shape, so patterns are visible later. The rows below are an illustration of the format, not real results:
| Question | Run | Your domain cited? | Which URL | Your brand recommended? | Domains cited instead |
|---|---|---|---|---|---|
| Which accounting software handles multi-currency invoices for a small agency? | 1 | No | - | No | a review site, a marketplace listing |
| Same question | 2 | Yes | /features/multi-currency/ | No | a review site |
Count breadth as well as volume. How many distinct questions cited your domain is a far more honest number than the total count of citations, because one answer citing several of your pages is not several times the reach.
Mention, citation and recommendation are three outcomes
The spreadsheet has separate columns for a reason. These three things happen independently, and merging any two of them hides the result you most need to see:
| Outcome | What it looks like | What it tells you |
|---|---|---|
| Mention | The answer text names your brand | The model knows you exist; says nothing about your pages |
| Citation | Your URL is marked as a source of the answer | Your page was used as evidence for this question |
| Recommendation | The answer suggests your product or service as the choice | You won the question, whichever pages were cited |
The combination that surprises people is a citation without a recommendation: your page is used as reference material while the answer recommends a competitor. It is specific, common and fixable, and a single combined score will never show it to you.
Reading a negative result
Once you have several runs per question, the pattern of the "no" tells you where to work:
| What you see | Likely cause | Where to look |
|---|---|---|
| Never cited; competitors cited | Access or extraction | Fetch the page as OAI-SearchBot; check the content is in the HTML without JavaScript; read the first two sentences under each heading out of context |
| Never cited; marketplaces, review sites or round-ups cited | The facts live on other people's pages | Check whether those pages carry your current facts, prices and availability |
| Cited, but a competitor is recommended | Corroboration | Third-party mentions, directory listings and comparison pages you do not own |
| Cited in some runs and not others | Normal variance | Nothing to fix; measure a rate across many questions and runs instead of judging single answers |
The second row surprises people most. Comparison and round-up pages match the shape of a buying question, and in commercial categories they often take the citation even when your own page states the same facts. A listing page carrying your facts with third-party framing can beat a perfect product page. For many brands the fastest win is getting those third-party pages right rather than rewriting their own. Third-party corroboration covers how to find them.
For the first row, the extraction test is simple: read the first two sentences under each heading with everything else hidden. Assistants lift sentences that answer a constraint, such as a dimension, a price condition, a return window or a compatibility. Copy written for brand tone gives them nothing to lift, however good it reads. Writing for extraction shows how to rewrite such passages.
The third row is a corroboration problem. A model needs the same fact from independent places before it states it confidently, and a claim that exists only on your own domain is the weakest evidence you can offer.
If you are building this check into a tool
Automating the check introduces its own failure modes, and they all make the result look better or worse than it is:
- Derive "cited" from the citation URLs, not from the answer text. Matching your domain against the actual source URLs is the only reliable test. Text that mentions your brand is a separate signal.
- Count marked sources, not the retrieval pool. An assistant typically retrieves a dozen or more pages and cites a few. Counting what was retrieved instead of what was marked overstated our own citation figures by two to four times before we fixed it.
- Discard markers that point at nothing. Models occasionally emit a marker that resolves to no source. Record one citation fewer rather than one that is not real.
- Keep cited and recommended in separate columns. Merging them flatters the report and hides the gap.
- Never turn a failed run into a zero. A request that timed out produced no evidence either way. Recording it as "not cited" drags the rate down and turns a normal month into an apparent decline.
Citation analysis explains how to turn the raw records into something you can act on.
Cited is not recommended, and cited is not traffic
A citation also does not mean a visit. Most answers are read without a click, so if you look for ChatGPT's use of your site only in your analytics, you will undercount it badly. Measure on the answer side: which questions cite you, how consistently, and what the answer says about you when it does. Analytics shows the small share of readers who clicked through; the answer log shows the rest.
If your pages answer questions that imply a current answer, state a real published and modified date on them. Undated pages, or pages whose modified date changes on every deploy, give an engine no honest signal of freshness, and the second case is a claim a crawler can check against the content.
A clean result is not a guarantee either. Passing all four checkpoints makes a citation possible; it does not make one certain, and nothing you change on your own site can promise a place in a particular answer. What the check gives you is a defensible picture of where you stand and a short list of what to fix first.
Sources
- OpenAI, Overview of OpenAI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User), checked 2026-09-09.
- OpenAI, OAI-SearchBot IP ranges.